Open Topics
If you are a Bachelor’s or Master’s student at TU Berlin and interested in writing your thesis on one of the following topics, please contact me via mail (see sidebar).
You should have a solid background in and have taken prior courses related to natural language processing and/or machine learning.
At the moment, I’m not in the position of supervising PhD students on my own, but I’m always happy to provide consultation on an informal basis!
Verbalization of layer functions Translating layer analyses “globally” (across a whole dataset) and “locally” (for single instances) into natural language
[1] MAPS (Elhelo & Geva, ACL 2025)
[2] PRISM (Kopf et al., NeurIPS 2025)
[3] LatentQA (Pan et al., 2024)
[4] Talking Heads (Merullo et al., NeurIPS 2024)
[5] Hou & Castanon (ICML 2023)
[6] Layer by Layer (Zhao et al., EMNLP 2024)
[7] Information Flow Routes (Ferrando & Voita, EMNLP 2024)
[8] Ikeda et al. (COLM 2025)
Contrastive attribution for readability-controlled generation
[1] Hsu et al. (GEM @ ACL 2025)
[2] Yin & Neubig (EMNLP 2022)
[3] Agrawal & Carpuat (TACL 2024)
[4] Barayan et al. (COLING 2025)
[5] Buçinca et al. (CHI 2025)
[6] RSA-Control (Wang & Demberg, EMNLP 2024)
Agentic auto-interpretability (with computational constraints) Designing small-scale LLM agents with self-testing and self-interpretability tools
[1] MAIA (Shaham et al., ICML 2024)
[2] CRITIC (Gou et al., ICLR 2024)
[3] Liu et al. (NAACL 2024)
[4] Kim et al. (2025)
[5] Ferrando et al. (2024)
Measuring the influence of verbatim-memorized content in training data on complex reasoning tasks
[1] Huang et al. (EMNLP 2024)
[2] Carlini et al. (ICLR 2023)
[3] Prabhakar et al. (EMNLP 2024 Findings)
[4] STIM (Li, Chen et al., 2025)
[5] Reason to Rote (Du et al., EMNLP 2025)
[6] Morris et al. (2025)
[7] ParaPO (Chen et al., COLM 2025)
[8] Survey on Memorization (Xiong et al., 2025)
Estimating the influence of LLM sycophancy on user interactions
[1] ELEPHANT (Cheng et al., 2025)
[2] Farm (Xu et al., 2024)
[3] The Siren Song of LLMs (Shi et al., 2025)
[4] Epistemic Alignment (Clark et al., COLM 2025)
[5] Don't Be Fooled (Spitzer et al., IJHCI 2025)
[6] IQA-EVAL (Li et al., NeurIPS 2024)
Text simplification of medical terminology
[1] FactPICO (Joseph et al., ACL 2024)
[2] README (Yao et al., EMNLP 2024 Findings)
[3] Two-Pronged Human Evaluation of ChatGPT Self-Correction in Radiology Report Simplification (Yang et al., ACL 2024 Findings)
[4] Effectiveness of ChatGPT in explaining complex medical reports to patients (Sun et al., 2024)
[5] Generative Artificial Intelligence to Transform Inpatient Discharge Summaries to Patient-Friendly Language and Format (Zaretsky et al., JAMA 2024)
[6] Fool Me Once? Contrasting Textual and Visual Explanations in a Clinical Decision-Support Setting (Kayser et al., EMNLP 2024)
Tracing biomedical knowledge in LLMs
[1] Large language models encode clinical knowledge (Singhal et al., Nature 2023)
[2] DILA: Dictionary Label Attention for Mechanistic Interpretability in High-dimensional Multi-label Medical Coding Prediction (Wu et al., PMLR 2025)
[3] Elucidating Mechanisms of Demographic Bias in LLMs for Healthcare (Ahsan et al., EMNLP 2025 Findings)